Large-scale underwater fish recognition via deep adversarial learning

Fish species recognition from images captured in underwater environments plays an essential role in many natural science studies, such as fish stock assessment, marine ecosystem analysis, and environmental research. However, the noisy nature of underwater images makes it difficult to train high-performance fish recognition models. This work presents a novel deep adversarial learning framework called AdvFish to train accurate deep neural networks fish recognition models from noisy large-scale underwater images. Unlike existing methods that rely on feature engineering or implicit machine learning techniques to mitigate the noise, AdvFish is a min-max bilevel adversarial optimization framework that trains the model on adversarially perturbed images via a proposed adaptive perturbation method. We show, on multiple benchmark datasets, that AdvFish holds a clear advantage over existing methods/models, especially on a noisy large-scale dataset. AdvFish is a generic learning framework that can help train better recognition models from extremely noisy images.

Automated summary

Fish species recognition from noisy large-scale underwater images is challenging. This work presents a novel deep adversarial learning framework, AdvFish, which outperforms existing methods/models on multiple benchmark datasets. AdvFish is a generic learning framework that can train better recognition models from extremely noisy images.